Saw y’all taking this test so I took this test. Some answers miss nuance though, it’s hard to pick sometimes, and “will AI” questions have no time horizon… Anyways 🤷🏻‍♂️ bambamramfan.github.io/ai-compass/

Man, that was a comeback

This World Cup is exposing me to the insane levels of absolutely weird commercials on US TV and I now understand why they made Netflix

I just joined @sifa.id :) sifa.id/p/danielesal...

I guess I need to uninstall claude now How did they even think this was acceptable behavior?

Good imo Will speed up processes already in fieri

Fable is getting out of prison

Still true even in the age of LLMs: The first 90% of the code accounts for the first 90% of the development time. The remaining 10% of the code accounts for the other 90% of the development time.

I don’t want to give my ID to use an LLM. So I’ll need to either hope for OpenAI to do things differently, hope for Mistral to release something, or wait for open-weight Chinese models I guess.

This would be nice How did we let things get so bad?

Just what I needed for a little project of mine... 😏

This was a pretty good weekend I need to spend more time afk

Found this in the public library in Vancouver, and I wish I could borrow it

One of my favorite car camping spots Sometimes one just needs to go outside a bit

I assumed most of you knew about/used this feed You should try it if you haven’t already!

Safety for Thee, Frontier Models for Me A critique of AI safety as export control, and why Europe needs local inference, open-weight models, and real autonomy from US-controlled AI. https://danielesalatti.it/en/blog/safety-for-thee-frontier-models-for-me

Interesting paper: "By exploiting the associativity of matrix multiplication, GMA avoids materializing the induced N×N affinity matrix and instead uses two responsibility matrices whose dominant activation storage scales as O(NK) rather than O(N2) for fixed K." arxiv.org/pdf/2606.18283

@unsloth.ai published quantized versions of GLM 5.2 but they haven’t announced it here yet huggingface.co/unsloth/GLM-...

I so wish I could run GLM 5.2 locally, even quantized and at 20 tok/s would be enough This thing is so good

Let's goooo! They opened it!

SpaceX acquired Cursor :/

I have thoughts about this AI/LLM copyright thing, especially compared to what happened in the early 2000s It’s nuanced and I think both sides are missing the point I need to spend some time and put my thoughts into words

Drove through Nevada (SLC to LV) These things look so much like they are out of a dystopian movie

Running Qwen 3.6 MoE MTP local has been a joy I know it's not SOTA, but it's **local** and it's **fast**

Interesting timing…is this due to the loss of users that the change caused? Pushback working? Or something else?

Who are the EU AI/LLM researchers on here? I want to follow them.

Don’t remember who said this, but the jailbreak was apparently asking fable to “fix all bugs you find in this code”, and then reverse engineer which ones were security vulns from the diffs (I’m guessing with other models). If that’s the case this is going to be hard to patch.

GLM 5.2 released and soon to come as Open Weight under MIT license digg.com/tech/ii9xibgn

Last week could have been designed in a lab to demonstrate how badly screwed we are if we rely on general-purpose intellectual technologies designed by an oligopoly of companies in one or two nations. Access restrictions — banned topics — effing sophons that silently degrade you.

What Europe should do right now: 1. Call all the European researchers working on AI and return them back with same salary (or they can stay but switch career). 2. Fill EU places having GPUs with money, and put those people there. 3. AI partnerships with China + India.

Is this going to be the crypto wars all over again? Someone go get Phil Zimmermann, stat

Looks like my Twitter account will be very short lived, given musk’s recent comments on EU stuff Why did I even think he’d finally decided to shut up? Lemme go and delete…

It’s kinda hilarious that they nerfed Fable so much it underperforms Codex 5.5 in some benchmarks agents-last-exam.org/leaderboard

Kimi 2.7 Code released, and MiniMax M3 is now open weights

I’m working on this! But not just cost savings, also for compliance and sovereignty Cost savings is only one angle of model routing

Distillation should be legally protected

Parallel LLMs?

Good article on Claude Fable 5 by @natolambert.bsky.social open.substack.com/pub/robotic/...

New Gemma, and it's *diffusion* text model blog.google/innovation-a...

I want to say a final thing about my Fable first reaction: I dedicated my life to programming and I'll use every innovation in the field, also to extract value and bring it to the local inference world, to Redis, and so forth. But:

My two main concerns properly expressed by someone who’s better with words than me

Yeah I don’t like this direction either Are open models going to be the only sensible way forward?

And we want to see more models!

"Silently degrading" on ML work effectively means I can't trust (or use) this model

Apple blaming the EU for their own faults... "Instead ​of trying to find ⁠a suitable compliance solution, Apple simply made a request to the European Commission to be exempted from their interoperability obligations under the DMA - and this for at least 18 months. ​That's not an option"

It looks like it’s the first model specifically trained to organize and delegate to other models

I ended up doing it and it's sooo annoying to dig through all the engagement bait

Oh, this is the deal breaker: it's available on all plans until Jun 22nd, then > On June 23, we’ll remove Fable 5 from those plans. Using it after that will require usage credits. If capacity allows, we’ll extend the included window.

Conveniently the cost is on a *logarithmic* scale

So the new Claude is going to be called Fable